Use of Scenario Ensembles for Deriving Seismic Risk

Use of scenario ensembles for deriving seismic risk

Tom R. Robinsona,1, Nicholas J. Rossera, Alexander L. Densmorea, Katie J. Ovena, Surya N. Shresthab, and Ramesh Guragainb

aDepartment of Geography, Durham University, Durham DH1 3LE, United Kingdom; and bNational Society of Earthquake Technology–Nepal, Kathmandu, Nepal

Edited by John Vidale, University of Southern California, and approved August 20, 2018 (received for review May 2, 2018) High death tolls from recent earthquakes show that seismic risk estimated recurrence intervals (13, 14). The resulting output is remains high globally. While there has been much focus on seismic an estimate of the likelihood of exceeding some value of ground hazard, large uncertainties associated with exposure and vulner- motion at a given location over a given period of time (e.g., a 2% ability have led to more limited analyses of the potential impacts chance of exceedance in 50 y). Thisisespeciallyusefulforde- of future earthquakes. We argue that as both exposure and termining appropriate seismic design codes for built infrastructure, vulnerability are reducible factors of risk, assessing their impor- allowing engineers to establish the maximum strength of shaking tance and variability allows for prioritization of the most effective that buildings are expected to witness during their design life (14). disaster risk-reduction (DRR) actions. We address this through Despite its sound basis, PSHA can be misunderstood, leading to earthquake ensemble modeling, using the example of Nepal. We implementations that attract criticism (15). This is especially true in model fatalities from 90 different scenario earthquakes and regions where past earthquake data are sparse (2, 11, 16–18), where establish whether impacts are specific to certain scenario earth- spurious probabilities can be generated (11). These criticisms have quakes or occur irrespective of the scenario. Our results show that proved controversial, however (19, 20), and several have been for most districts in Nepal impacts are not specific to the particular largely rejected (21). Nevertheless, in regions with limited in- characteristics of a single earthquake, and that total modeled formation on future earthquake probabilities different applica- impacts are skewed toward the minimum estimate. These results tions of PSHA can result in widely differing hazard and risk suggest that planning for the worst-case scenario in Nepal may estimates, such as recent efforts in Nepal (22). place an unnecessarily large burden on the limited resources avail- DSHA focuses on the use of scenarios of individual or small able for DRR. We also show that the most at-risk districts are predominantly in rural western Nepal, with ∼9.5 million Nepalis numbers of earthquakes, typically considering either the maximum EARTH, ATMOSPHERIC, AND PLANETARY SCIENCES inhabiting districts with higher seismic risk than Kathmandu. Our credible event or the worst-case event that could occur on known proposed approach provides a holistic consideration of seismic risk active or potentially active faults (14, 23). Shaking from the for informing contingency planning and allows the relative impor- resulting scenario(s) is derived from attenuation relationships us- tance of the reducible components of risk (exposure and vulnera- ing different likelihoods of exceedance (14). The resulting output bility) to be estimated, highlighting factors that can be targeted shows the strength and extent of shaking expected from the most effectively. We propose this approach for informing contin- maximum credible or worst-case earthquake with a given like- gency planning, especially in locations where information on the lihood of exceedance, providing an upper limit for planning. likelihood of future earthquakes is inadequate. This approach also has notable limitations, however, such as (i) a focus on one or a small number of events, (ii) difficulty in scenario ensembles | seismic risk | contingency planning | earthquakes | accurately determining the maximum credible event, and (iii)a hazard and risk weak statistical basis for estimates of uncertainty (19, 20, 24).

espite global efforts to reduce seismic risk, earthquakes re- Significance Dmain one of the deadliest natural hazards worldwide (1). Much of the scientific interest in reducing seismic risk, which is a High death tolls from recent earthquakes have highlighted the function of hazard, exposure, and vulnerability, has focused on need to better identify ways to effectively reduce seismic risk. better understanding of seismic hazard, with a particular focus We address this need by developing a new earthquake sce- on refining estimates of recurrence times and probabilities of nario ensemble approach. We model impacts from multiple exceeding given levels of ground motion (2, 3). While hazard different earthquake scenarios, identifying impacts that are assessment is a prerequisite for calculating risk, available data on common to multiple scenarios. This method allows us to esti- exposure and functions that model fragility often introduce sig- mate whether particular impacts are specific to certain earth- nificant uncertainties. Furthermore, full risk calculations require quakes or occur irrespective of the location or magnitude of a holistic analysis of losses, including fatalities, injuries, and fi- the next earthquake. Our method provides contingency plan- nancial, infrastructure, property, and indirect losses, so deriving ners with critical information on the likelihood, and probable absolute risk is often intractable. Consequently, while there have scale, of impacts in future earthquakes, especially in situations been several notable advances in the computation of earthquake where robust information on the likelihood of future earthquakes is incomplete, allowing disaster risk-reduction efforts risk and probable loss at national and global levels (4–10), these to focus on minimizing such effects and reducing seismic risk. have tended to focus on data-rich regions, such as California

(11). Despite these efforts, the high death tolls in many recent Author contributions: T.R.R. and N.J.R. designed research; T.R.R. performed research; large earthquakes demonstrate that earthquake risk remains T.R.R., N.J.R., A.L.D., K.J.O., S.N.S., and R.G. analyzed data; and T.R.R., N.J.R., A.L.D., and high globally, and in data-poor regions such as the Himalaya may K.J.O. wrote the paper. even be increasing as growth in population exposure and vul- The authors declare no conflict of interest. nerability outpaces the rate of improvement in understanding of This article is a PNAS Direct Submission. seismic hazard (1, 11, 12). This open access article is distributed under Creative Commons Attribution License 4.0 (CC BY). The two most common approaches to seismic hazard analysis 1To whom correspondence should be addressed. Email: [email protected]. (SHA) are probabilistic (PSHA) or deterministic (DSHA). This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10. PSHA is a widely used method that identifies all known possible 1073/pnas.1807433115/-/DCSupplemental. earthquakes that may affect a given site and characterizes their

www.pnas.org/cgi/doi/10.1073/pnas.1807433115 PNAS Latest Articles | 1of10 Downloaded by guest on September 24, 2021 Irrespective of the approach used, the outputs of both are 70E 80E 90E 100E 2005 arguably not tailored for contingency planning, where defining 1555 risk in terms of the potential consequences of the next future 1905 China 1430 earthquake is the priority concern. Contingency planning oper- Pak. 1344 30N 1803 1833 MHT 30N ates on two levels: first through planning for times of disaster and 1505 2015 1100 second for disaster risk reduction (DRR) (25, 26). Effective 1255 1950 planning requires both estimation of the likelihood and scale of 1934 1714 future earthquake impacts and understanding of those that are 25N 1897 25N specific to a single earthquake scenario or that could occur in India Ban. many different earthquakes. Likewise, effective contingency planning requires that we can determine the locations where impacts are 20N Myanmar 20N most likely to occur, along with the average and worst-case impacts 70E 80E 90E 100E for all locations, so that both emergency relief and preevent DRR > activities can be prioritized. Thus, for those tasked with managing Fig. 1. Earthquake history of the Himalayan arc. Numerous large (Mw 7.0) earthquakes have been recorded along the MHT system over the last 1,000 y, earthquake risk, moving beyond probabilities of shaking to proba- with little evidence that the largest ruptures are confined to any specific bilities of consequences of future earthquakes is essential (25, 27). segment. Polygons show known or inferred rupture extents with associated

Addressing such complex questions about future events reso- calendar dates and colors represent magnitudes (green, Mw 7.0–8.0; orange, nates with the challenges faced by climate and meteorological Mw 8.0–8.5; red, Mw 8.5+). Dashed box shows location of Nepal. Red lines modelers attempting to generate future climate and weather show active faults from Taylor and Yin (82). Ban, Bangladesh; Pak, Pakistan. scenarios. They address this through the use of ensembles of models, which consist of suites of scenarios of future climate or “ ” weather events based on different conditions and model reali- Thus, we focus on relative risk between scenario outcomes, which we argue is invaluable for earthquake contingency planning. Uniform proba- zations (28–34). The outputs from all scenarios are then aggre- bilities will also overemphasize the contribution from the largest-magnitude gated to identify common elements that are more likely to be events, as well as those on upper-plate faults. Conversely and importantly, realistic representations of future events. Here, we propose a uniform weighting allows a focus on the role of exposure and vulnerability similar approach for the assessment of seismic risk, to derive in producing risk and impacts. This is crucial for contingency planning and greater clarity on the potential impacts of future earthquakes. DRR, because while earthquake hazard is irreducible, both exposure and We establish an ensemble of earthquake scenarios, with each vulnerability to earthquakes can be reduced. individual scenario containing empirically derived estimates of While the recurrence interval for each of our scenarios is unlikely to be the associated impacts. We then average and compare conse- uniform, previous work has suggested that earthquakes of all magnitudes quences from all scenarios in the ensemble to examine the on the Main Himalayan Thrust (MHT) have ∼500-y recurrence intervals, and > emergent impacts, focusing on those that are common to mul- that major [moment magnitude (Mw) 7] earthquakes can be followed by > tiple scenarios. Our approach is not intended to supersede either great (Mw 8) earthquakes in the same location sooner than plate conver- gence rates would suggest possible (36). Such observations may explain the PSHA or DSHA, as no individual analysis is suitable for all relatively short intervals between the 1833 (M 7.8), 1934 (M 8.2), and 2015 intended tasks (14). Instead, we propose the approach as a w w (Mw 7.8) earthquakes in central and eastern Nepal (Fig. 1). Others, however, complementary tool for the assessment of seismic risk with the have suggested that recurrence intervals for the largest magnitude events specific aim of informing earthquake contingency planning. We on the MHT may be on the order of 1,000 y (45). Importantly, however, this concentrate here on providing the median and maximum impact highlights that at present we remain unable to assign meaningful re- estimates, the number of impact-inducing scenarios, the speci- currence intervals beyond uniform. ficity of impacts to individual scenarios, and exceedance proba- In each scenario, we combine estimates of ground shaking with population bilities for impacts. We demonstrate the approach using the case and building exposure data taken from the most recently available (2011) of earthquake-induced fatalities in Nepal. Earthquake hazard in national population census of Nepal (46) and use previously published, Nepal is relatively poorly constrained, leading to often widely empirical building fragility curves to estimate resulting impacts. We calculate differing hazard maps (22), but is thought to be among the fatalities by Village Development Committee (VDC), which was the third- level administrative division in Nepal up to 2017 and is consistent with the highest globally (35–38). Population exposure and vulnerability 2011 census data. We then aggregate fatalities and fatality statistics (fre- to earthquakes is also high (39, 40), and previous earthquake – quency, median, maximum, and specificity) across the 75 districts and five impacts have been substantial (41 44), yet impact estimates for development regions, which comprised the pre-2017 first- and second-level future earthquakes are limited (42). While we focus on fatalities, administrative divisions. We focus on fatalities as a single measure of impact, other forms of loss (injuries and financial losses) could also be but other loss measures such as injuries, financial losses, or property losses explored in this manner. could equally be calculated. Because we focus on relative risk, the numbers of fatalities discussed below are only indicative of expected impacts, and Materials and Methods they are not intended as absolute estimates of likely fatalities in Nepal. Fi- Method Overview. We adapt the ensemble approach used in climate and nally, to provide information tailored to earthquake response planners, we meteorological modeling for the purposes of estimating the consequences of also consider social vulnerability, which has been shown to be a significant future earthquakes. We model the losses associated with 30 different predictor of earthquake impacts and losses (47). We combine our physical earthquakes that are large enough to cause substantial damage in Nepal at vulnerability metrics with two examples of social vulnerability metrics three different times of day to give 90 scenarios, based on our current un- employed as proxies for disaster vulnerability: the Human Development In- derstanding of active fault locations and potential future earthquakes. The dex (HDI) (48) and a remoteness index that reflects the need for and ease of sample of scenarios is chosen based on current understanding of historic providing logistical assistance, to estimate total relative seismic risk for each earthquakes (Fig. 1) and fault slip rates to give a suite of geologically diverse district of Nepal. prototypical scenarios and is large enough that the statistical properties of the results can give some useful insight into the possible consequences of Modeled Earthquake Scenarios (Hazard). We choose an ensemble of 30 large these earthquakes. While each of the modeled earthquakes is plausible, the (Mw >7) earthquake scenarios based on historical records and paleoseismic exact probability of each remains unknown. Instead, each scenario is evidence (Fig. 1), assuming that previously documented earthquakes are assigned a uniform probability and weighting in the ensemble. While this representative of potential future earthquakes at decadal-to-centennial approach avoids issues associated with selection of weights based on poorly time scales (49). Note that this approach cannot account for unanticipated constrained recurrence intervals, it has important consequences for our re- events such as fault linkage or simultaneous rupture of multiple faults (e.g., sults. First, a uniform weighting precludes the ability to discuss “absolute ref. 38). For known or inferred active faults without historical evidence of risk,” because the hazard calculations do not include absolute probabilities. earthquakes, geologic data on long-term slip rates and displacement styles,

2of10 | www.pnas.org/cgi/doi/10.1073/pnas.1807433115 Robinson et al. Downloaded by guest on September 24, 2021 along with fault dimensions and empirical scaling relationships (50), were MHT_8.6_Utk MHT_8.6_FMW MHT_8.6_MWC MHT_8.6_WCE used to estimate plausible scenario earthquakes.

In the last 1,000 y, at least 15 Mw >7.5 earthquakes have been recorded along the Himalayan arc (Fig. 1) (36–38, 44, 51). The majority of these are associated with the MHT; however, spatial variations in rheology and geometry can limit rupture extent, giving rise to various prototypical forms of MHT_8.6_Sik MHT_8.3_Utk MHT_8.3_Far MHT_8.3_Mid MHT earthquake (36, 51). These include (i) giant ruptures, such as the 1950 Assam and 1505 western Nepal earthquakes, that initiate near the brittle– ductile transition and rupture to the surface, have lengths >200 km, and > – have Mw 8.5 (36, 37); (ii) great ruptures, such as the 1934 Nepal Bihar MHT_8.3_Wst MHT_8.3_Cen MHT_8.3_Est MHT_8.3_Sik earthquake, that are similar to giant ruptures but do not necessarily reach

the surface and have Mw 8.0–8.5 (35, 44); and partial ruptures, like the 2015 Gorkha event, that rupture either the (iii) lower or (iv) upper ramp of the MHT and have M 7.0–8.0, with larger magnitudes anticipated on the lower w MHT_7.8_Utk KKM_7.8_Chn MHT_7.8_Far MHT_7.8_Mid ramp (52) (Fig. 2). Paleoseismic evidence of great-to-giant earthquakes on the MHT in ca. 1100, 1255, and 1344 (Fig. 1) suggests that earthquakes on this fault are not constrained to individual segments within Nepal and can occur on any section of the MHT throughout the Himalayan arc (36, 44, 53). As well as the MHT, numerous other faults within or close to Nepal have WFS_7.8_Mid MHT_7.8_Wst MHT_7.8_Cen MHT_7.8_Est

previously sustained, or are capable of sustaining, Mw 7+ earthquakes. The largest is the Karakorum Fault, which hosted a Mw ∼7.5 earthquake in 1895

MHT_7.8_Sik MHT_7.3_Utk MHT_7.3_Far MHT_7.3_Mid

A KKM SN MCT TKK MCT

Mw 8.5+ WFS_7.3_Wst MHT_7.3_Wst TKK_7.3_Wst MHT_7.3_Cen MFT MFT B KKM MCT TKK EARTH, ATMOSPHERIC,

WFS MHT_7.3_Est MHT_7.3_Sik AND PLANETARY SCIENCES MFT Mw 8.0-8.5 0.0 0.4 0.8 1.2 1.6 2.0 C KKM Peak Ground Acceleration MCT TKK MCT PGA (g) WFS Mw 7.5-8.0 MFT MFT Fig. 3. Earthquake scenario ensemble. Modeled ground shaking in terms of KKM PGA with 50% probability of exceedance for the 30 scenario earthquakes in D the ensemble. Note that shaking values are only shown for locations within MCT TKK MCT Nepal. Scenario codes are given in the format fault_magnitude_location. WFS Mw 7.0-7.5 MFT MFT Cen, Central Region; Chn, China; Est, East Region; Far, Far-West Region; FMW, Far-West, Mid-West, and West Regions; KKM, Karakorum Fault; Mid, E KKM Mid-West Region; MWC, Mid-West, West, and Central Regions; Sik, Sikkim MCT MCT (northeast India); TKK, Thakkhola graben; Utk, Uttarakhand (northwest In- WFS Mw 7.0-7.5 dia); WCE, West, Central, and East Regions. MFT MFT F KKM MCT TKK MCT (54) and is capable of Mw 8.0 events (55). In western Nepal, a set of faults known as the Western Fault System (WFS) partition motion between the WFS Mw 7.0-7.5 MFT MFT MHT and the Karakorum Fault. Quaternary offsets associated with these G KKM faults suggest repeated earthquakes since the last glacial advance (56) with evidence of possibly two Mw 7+ earthquakes between AD 1165 and 1400 MCT TKK MCT (57). Extension in the southern Tibetan Plateau is accommodated on a series WFS Mw 7.5-8.0 – MFT MFT of north south-striking normal faults, of which the largest, most active, and closest to Nepal are the faults bounding the Thakkhola graben. These have

H historically sustained Mw 6.2–6.4 earthquakes but are likely capable of Mw + MCT TKK MCT 7 events (58, 59). We therefore consider eight different prototypical scenarios for Mw 7+ WFS Mw 7.5-8.0 MFT MFT earthquakes in Nepal (Fig. 2). Earthquakes on upper-plate faults such as the WFS and the Thakkola graben are restricted in their location, whereas those Fig. 2. Map and simplified north–south cross-sectional views of the eight occurring on the MHT are allowed to occur at multiple locations along strike. prototypical scenario earthquakes in our ensemble. Red-outlined boxes For the MHT, we assign earthquake magnitudes at the center of the pub-

(Left) show the surface projection of the assumed failure planes. Thick red lished ranges, comprising (i) giant earthquakes with Mw 8.6, (ii) great lines in (Right) show the down-dip extents of fault rupture, while dashed earthquakes with Mw 8.3, (iii) blind lower-ramp earthquakes with Mw 7.8, lines show possible simultaneous/alternative rupture scenarios. (A) giant (Mw and (iv) upper-ramp earthquakes with Mw 7.3. We model these earthquakes 8.5+) earthquakes on the MHT such as the 1505 western Nepal event; (B) as occurring between Uttarakhand on Nepal’s western border and Sikkim to

great (Mw 8.0–8.5) earthquakes on the MHT such as the 1934 Nepal–Bihar the east, incrementally shifting each rupture patch to produce adjacent event; (C and D)Mw 7.0–8.0 ruptures of the lower or upper ramp of the scenarios that span and extend beyond Nepal to avoid edge effects. In total, MHT, similar to the 2015 Gorkha event; (E)Mw 7.0–7.5 ruptures of normal we consider five giant scenario earthquakes and seven of each of the great, faults in southern Tibet, such as those bounding the Thakkhola graben (note upper ramp, and blind lower ramp scenario earthquakes (Fig. 3). For the

that rupture is not shown in cross-section); (F)Mw 7.0–7.5 ruptures of the upper-plate faults, we consider events at the upper end of the likely mag- southern portion of the WFS; (G)Mw 7.5–8.0 ruptures of the northern por- nitude range: (v)aMw 7.8 earthquake on the southern part of the Kar- tion of the WFS; (H)Mw 7.5–8.0 ruptures of the Karakorum Fault. KKM, akorum Fault, (vi)aMw 7.8 event on the northern part of the WFS, (vii)aMw Karakorum Fault; MCT, Main Central Thrust; MFT, Main Frontal Thrust; TKK, 7.3 event on the southern part of the WFS, and (viii)aMw 7.3 earthquake in Thakkhola graben. Data from ref. 51. the Thakkhola graben (Fig. 3).

Robinson et al. PNAS Latest Articles | 3of10 Downloaded by guest on September 24, 2021 While exposure as a function of both daily and seasonal variations in A Population (1000s) building occupancy is still poorly understood, we account for some temporal differences by deriving building occupancy rates for three different earthquake occurrence times: (i) night, (ii) day (working), and (iii) day (nonworking). We distinguish between urban and rural VDCs by assuming that urban locations have higher occupancy on working days than rural locations, and vice versa. Building occupancy rates (Table 1) are derived in consultation with international humanitarian partners based in Nepal and are subject to a first-order calibration through retrospective fatality modeling of the 2015 Gorkha earthquake (SI Appendix). We note, however, that these assumptions and associated uncertainties can be large and so represent a consid- 0 1 2.5 5 7.5 10 25 50 erable gap in current knowledge. B Adobe (100s) C Bamboo/timber (100s) Vulnerability. We derive total fatality estimates for each scenario by considering the vulnerability of each building typology to seismic shaking, combining locally (64) and globally derived (10, 65) building fragility data where necessary. Based on the work of the Global Earthquake Model–Earthquake Consequence Database (GEM-ECD) (65), we assume that shaking-derived fatalities are lim- 0 1 2.5 5 10 25 50 100 0 1 2.5 5 10 25 50 100 ited to collapsed buildings, which correspond to a subsection of the “Complete D Reinforced concrete (100s) E Brick & concrete (flexible) (100s) Damage” state described in HAZUS (10). We therefore calculate the number of buildings suffering complete damage using the relevant fragility curves, before estimating the proportion that collapse based on probabilities from the GEM-ECD (Table 2). For adobe, brick and concrete (flexible flooring), brick and concrete (rigid flooring), brick with mud mortar, and stone with mud mortar buildings, we 0 1 2.5 5 10 25 50 100 0 1 2.5 5 10 25 50 100 use available Nepal-specific fragility curves (Fig. 5) from Guragain (64). These F Brick & concrete (rigid) (100s) G Brick w/ mud mortar (100s) predominantly masonry buildings are most prevalent throughout Nepal, accounting for 65% of the total and almost all buildings in rural regions (Fig. 4). For nonengineered reinforced concrete and bamboo/timber buildings, no Nepal-specific fragility curves are available and thus we rely on fragility curves from HAZUS (10), using the curves corresponding to building types C3M (concrete frame with unreinforced masonry infill, midrise, low code) 0 1 2.5 5 10 25 50 100 0 1 2.5 5 10 25 50 100 and W1 (wood, light frame, low code), respectively (Fig. 5). We note that these curves were initially developed for the United States and may not be H Stone w/ mud mortar (100s) I ADB (1%) applicable to Nepal. Despite this, the curve for reinforced concrete structures WDN (25%) suggests a worse performance than found in recent empirical analysis of SMM (35%) building performance during the 2015 earthquake (66) and so is likely to be conservative. NRC (10%) Finally, to estimate total seismic risk by district we combine fatality sta- 0 1 2.5 5 10 25 50 100 BMM (12%) BCF (7%) tistics from the ensemble with two social vulnerability measures: remoteness BCR (12%) and human development. Remoteness is a semiquantitative measure of accessibility for each VDC developed by the US Agency for International De- Fig. 4. Population and building exposure in Nepal. Total population and velopment and scored out of 10 (1 = most accessible; 10 = least accessible). It number of residential buildings by construction type within each VDC in includes factors such as the distance to roads, available transportation Nepal according to the National Population and Housing Census (2011). (A) methods, and distance from key services. We use remoteness scores (67), population; (B) adobe buildings; (C) bamboo/timber buildings; (D) non- averaged across all VDCs in a district and weighted by population, as a engineered reinforced concrete buildings; (E) brick and concrete (flexible measure of predisaster accessibility. In the context of contingency planning, flooring) buildings; (F) brick and concrete (rigid flooring) buildings; (G) brick this measure is used as a proxy for the likely scale and speed of postdisaster with mud mortar buildings; (H) stone with mud mortar buildings; and (I)pie aid delivery, and by inference, an indicator of high levels of compounded chart showing the percentage of each building type. ADB, adobe; BCF, brick postdisaster vulnerability. It can also be considered as a measure of the likely and concrete (flexible flooring); BCR, brick and concrete (rigid flooring); need for postdisaster assistance, as remote rural communities have been BMM, brick with mud mortar; NRC, nonengineered reinforced concrete; shown to be more likely to require assistance than more accessible urban SMM, stone with mud mortar; WDN, bamboo/timber. communities (48). HDI is a summary measure of life expectancy, education, and standard of living, among other factors, and is scored out of 1, where 1 is most developed and 0 is least developed. We use the 2014 HDI scores for We model the shaking from each of these events in terms of peak ground each district of Nepal (68) as a proxy for human vulnerability to earthquakes, acceleration (PGA, in units of meters per second2) with OpenSHA (60), using with lower scoring districts considered more vulnerable. HDI has previously the ground motion prediction equations of Abrahamson and Silva (61), been investigated as an indicator for disaster risk, with higher HDI scores exceedance probabilities of 50%, and shallow shear wave velocity (Vs30) generally associated with lower average losses (48, 69). While both re- derived from topographic slope (62, 63). moteness and HDI have some direct relevance to social vulnerability, these measures are indicative rather than definitive and are not intended to ex- haustively capture all dimensions of social vulnerability to disasters. A more Exposure. We use the National Population and Housing Census 2011 for Nepal to assess the exposure of population and buildings (Fig. 4) to each scenario in the ensemble at the VDC level, the smallest pre-2017 administrative division Table 1. Building occupancy rates for which data are available. In the absence of alternative more reliable data, we do not disaggregate by gender or age. The census contains the Building occupancy number of residential buildings per VDC with different types of foundation, roof, and wall construction. Using this information, we classify residential Time of day Urban, % Rural, % buildings into seven different generic typologies: (i) adobe, (ii) bamboo/ Night 99 99 timber, (iii) brick and concrete (flexible flooring), (iv) brick and concrete Day (working) 70 40 (rigid flooring), (v) nonengineered reinforced concrete, (vi) brick with mud Day (nonworking) 40 70 mortar, and (vii) stone with mud mortar (Fig. 4). We estimate individual building occupancy by assuming a uniform distribution of people. Shaking Assumed residential building occupancy rates for urban and rural VDCs exposure for each scenario is derived using the mean modeled PGA per VDC. for three different times of day.

4of10 | www.pnas.org/cgi/doi/10.1073/pnas.1807433115 Robinson et al. Downloaded by guest on September 24, 2021 Table 2. Building collapse and fatality rates Building type Collapse probability, % Fatality rate, %

Adobe 15.0 5.0 Bamboo/timber 3.0 0.5 Brick and concrete (flexible) 15.0 5.0 Brick and concrete (rigid) 15.0 15.0 Nonengineered reinforced concrete 13.0 10.0 Brick with mud mortar 15.0 5.0 Stone with mud mortar 15.0 5.0

Collapse probabilities and fatality rates for different building types in Nepal derived from global empirical relationships and taken from GEM-ECD (65). Collapse probabilities apply only to buildings suffering “complete damage” as defined by HAZUS (10) and calculated from the respective fragility curves (Fig. 5).

definitive discussion of social vulnerability to natural hazards specific to fatalities are equivalent to those recorded in 2015, suggesting that Nepal is provided by Gautam (40). the Gorkha earthquake was close to a worst case for those districts.

Results Specificity of Impacts. Understanding how the impacts might vary Planning for Disaster. Because the exact nature of the next earth- under different earthquake scenarios is as important to contin- quake to occur is unknowable, we use our ensemble to estimate the gency planners as the median and worst-case impacts. If all relative scale of fatalities in the next earthquake, irrespective of its scenarios in the ensemble result in similar impacts in a district, nature, by assessing the frequency distribution of total earthquake then the district can be considered to have low hazard specificity. fatalities for all scenarios (Fig. 6). We find that over 70% of modeled Alternatively, if impacts are highly variable across the ensemble, ∼ scenarios result in more than the 9,000 fatalities experienced in the then a district has high hazard specificity, as the impacts are in- 2015 Gorkha earthquake (70), while 16% exceed ∼50,000 fatalities, ∼ trinsic to a precise scenario and so there is more uncertainty about and 2% exceed 100,000 fatalities. Based on our assumptions about what could happen in the next event. For contingency planning, building occupancy rates, there is a substantial increase in risk for

low specificity is preferable, even when associated with large im- EARTH, ATMOSPHERIC,

nighttime compared with daytime earthquakes. At night, 50% of AND PLANETARY SCIENCES pacts, as planners can be confident of the scale of impacts to be scenarios exceed ∼23,000 fatalities and 5% exceed ∼125,000 fatali- expected. For high-specificity locations, impacts are intimately ties, compared with ∼10,000 fatalities and ∼65,000 fatalities, re- linked to whichever earthquake occurs, but as this cannot be an- spectively, for daytime earthquakes (Fig. 6). Earthquakes in the Central Region incur the greatest losses, with 50% of scenarios ex- ticipated, specificity could inform planning decisions. ceeding ∼60,000 fatalities and 5% exceeding ∼144,000 fatalities, To calculate specificity, we determine the frequency distribu- compared with ∼11,000 fatalities and ∼54,000 fatalities, respectively, tion of impacts by district with respect to the corresponding for earthquakes in the Far-West Region. Only the M 8.6 scenarios worst-case scenario. The distribution is used to obtain the per- w centage of scenarios with fatalities exceeding a given fraction of generate in excess of ∼100,000 fatalities, while no Mw 7.3 scenario results in >50,000 fatalities. the worst case (Fig. 7). Calculating the area under the curve (AUC) indicates how losses are skewed toward either the mini- Risk Metrics. mum (AUC → 0), worst-case (AUC → 1), or are evenly dis- Fatality exceedance probabilities. We estimate the relative scale of tributed (AUC ∼ 0.5). (Fig. 7). Specificity is considered to be fatalities by district from the frequency distribution output from highest when AUC = 0.5 and reduces as AUC tends to 0 or 1. the entire ensemble (Fig. 7). A total of 72% of scenarios result in All districts have an AUC between 0 and 0.53, showing that fatalities in Kathmandu, the largest percentage of fatal scenarios impacts are either evenly distributed or skewed toward the for any district (Figs. 7 and 8). Districts in the East Region have the fewest number of fatal scenarios, typically <40% (Figs. 7 and 8). While this may appear to be an edge effect, the impacts of 100% scenarios occurring across the eastern border in Sikkim were included in the ensemble, and a similar result is not seen in the 80% Far-West Region related to the high proportion of timber/

bamboo buildings (Fig. 4). Importantly, as all districts have one 60% or more fatalities in at least one-third of the scenarios, seismic risk is high for the whole country.

damage 40% Median and worst-case fatalities. Median modeled fatalities are highest in Kavrepalanchok and the majority of the largest mod- 20% eled fatality totals are in the West and Central Regions (Fig. 8). Probability of complete Generally, districts that border China have the lowest median 0% fatalities, although notably some heavily populated districts in the 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 south also have low median fatalities. In Gorkha, Dhading, Lalitpur, PGA (g) and Nuwakot, the median fatalities are equivalent to those experi- Flexible brick & concrete Rigid brick & concrete Brick & mud / Adobe enced in the 2015 earthquake, suggesting that, in this sense, the Stone & mud Reinforced concrete Timber frame / Bamboo 2015 earthquake was a “typical” event in these districts. Maximum fatality estimates broadly correlate with the pop- Fig. 5. Residential building vulnerability. Empirically derived fragility curves ulation distribution (Fig. 4), with the three Kathmandu Valley for complete damage (i.e., the structure has collapsed or is in imminent danger of collapse) for different residential building types in Nepal from districts (Kathmandu, Lalitpur, and Bhaktapur) and the majority of Guragain (64) and HAZUS (10). Curves for nonengineered reinforced con- districts in the south having the largest worst-case fatalities (Figs. 7 crete and bamboo/timber buildings correspond to building types C3M and 8). Kathmandu has the largest worst-case fatalities at >24,000. (concrete frame with unreinforced masonry infill, midrise, low code) and W1 Notably, in Rasuwa and Sindhupalchok the maximum modeled (wood, light frame, low code) in HAZUS (10), respectively.

Robinson et al. PNAS Latest Articles | 5of10 Downloaded by guest on September 24, 2021 the worst case are likely to be rare. For the remaining districts, A 100% planning should focus on the worst-case impacts as fatalities are variable and dependent on the precise scenario that occurs. 80% Prioritization for Risk Reduction. With finite resources available for 60% risk-reduction efforts, contingency planning requires an objective approach to prioritize DRR efforts toward locations that are 40% exceeding most at-risk. To help inform this, we estimate the total relative seismic risk for all districts in Nepal by combining the probability 20%

Percentage of scenarios of fatalities, the median and maximum fatalities, and the speci-

0% ficity of fatalities with remoteness and HDI. We give each district 0 25 50 75 100 125 150 a normalized score out of 1 for all six risk metrics, such that the Fatalities (thousands) district considered most at risk (i.e., with the lowest value of HDI All scenarios Night Day (not working) Day (working) and the highest value for all other metrics) scores 1, and then simply sum for all of the metric scores assuming a uniform B weighting. We recognize that others may see value in alternate 100%

80% A 100% 100% 60% 80% 50% 40% exceeding 60% 0% 0 10,000 20,000 20% 40% Percentage of scenarios exceeding 0% 0 25 50 75 100 125 150 20% Fatalities (thousands) Percentage of scenarios 0% All scenarios Far-West Mid-West West Central East 1001 10 10 2310 10 410 5 C Fatalities 100% Bhaktapur Chitwan Kaski Kathmandu Lalitpur Solukhumbu

80% B 100% AUC→1

60% 80% AUC=0.5 AUC→0 40% 60% exceeding

20% 40% exceeding Percentage of scenarios

0% 20%

0 25 50 75 100 125 150 Percentage of scenarios Fatalities (thousands) 0% All scenarios Mw 8.6 Mw 8.3 Mw 7.8 Mw 7.3 0% 20% 40% 60% 80% 100% Percentage of worst-case fatalities Fig. 6. Exceedance probabilities for fatalities. Probabilities are derived from the frequency distribution of scenarios in the entire ensemble compared Bhaktapur Chitwan Kaski Kathmandu Lalitpur Solukhumbu with different scenario subsets: (A) time of day, (B) location of scenario Fig. 7. Fatality exceedance probability (A) and impact specificity (B) for all earthquake (based on pre-2017 development regions), and (C) earthquake 75 districts of Nepal. (A) Percentage of scenarios in the ensemble with fa- magnitude. The dashed gray line shows the number of fatalities recorded in talities exceeding given values for each district. Inset shows same data on a the 2015 Gorkha earthquake (70). linear scale. (B) Specificity of impacts in terms of their variability based on all scenarios causing >0 fatalities, normalized with respect to the worst-case scenario for each district. (Inset) Schematic definition of the data: Lines minimum. Crucially, no district has impacts skewed toward the with convex-up curvature (black) show that the majority of impacts are close worst case (Fig. 7). Worst-case impacts occur in very few sce- to the maximum (area under the curve, AUC, approaching 1), while lines narios, and the large majority of impacts are far less than the with concave-up curvature (green) show that the majority of impacts are maximum. For example, in Kathmandu 75% of fatality-inducing close to the minimum (AUC approaching 0). Both represent low specificity as scenarios result in fatalities that are <15% of the worst case. impacts show little variability with different scenarios. Linear distributions Importantly, there is large variation in specificity across Nepal: (red) show that impacts are evenly distributed (AUC ∼ 0.5) and thus repre- high-specificity districts are mostly clustered in the East Region, sent high specificity. In both panels, six key districts are highlighted: Kath- while low-specificity districts are along the southern border (Fig. mandu (red), Bhaktapur (green), and Lalitpur (purple) comprise the Kathmandu Valley and Nepal’s largest urban area; Kaski (blue) and Chitwan 8). For 55 of the 75 districts in Nepal, at least two-thirds of ’ < (yellow) host two of Nepal s other largest cities (Pokhara and Bharatpur, modeled scenarios result in impacts that are 50% of the worst respectively) and are popular tourist destinations; Solukhumbu (black) is case (Fig. 7). This suggests that contingency planning for these home to Mt. Everest and the Everest Base Camp trek, which is one of the districts should focus on median losses, as impacts approaching most popular treks in Nepal. All other districts are shown in gray.

6of10 | www.pnas.org/cgi/doi/10.1073/pnas.1807433115 Robinson et al. Downloaded by guest on September 24, 2021 AB of individual districts, but a similar overall pattern of higher risk scores in the west and a middle-to-low score for Kathmandu generally remains (SI Appendix, Fig. S3). Discussion The intention of this study is to outline an approach to the assessment of seismic risk that focuses on the importance of the reducible components of risk, namely exposure and vulnerability. C D We argue that this is critical for identifying and prioritizing the most pressing risk-reduction activities and the most at-risk locations at a national level. We do not intend for what we propose to supersede either PSHA or DSHA, but instead to complement them by specifically addressing the needs of contingency planners. It is therefore important to highlight the limitations of our ensemble approach and possibilities for further research. E F First, it is important to consider whether an ensemble can account for the full range of potential future earthquakes. We consider only a small number (8) of prototypical scenario earthquakes, although we allow their locations to vary. It is not clear how our results depend on the number of scenarios that are included in our ensemble, although in future this could be tested. Small changes in earthquake magnitude (∼0.1–0.2) compared with the larger steps between scenarios included here are un- G likely to affect our results, because ground motion saturation occurs at Mw 7.3–7.5, beyond which point the main factor controlling shaking strength is distance to the fault. Small increases (or decreases) in magnitude are also unlikely to require signifi-

cant changes in fault dimensions and therefore will not signifi- EARTH, ATMOSPHERIC, cantly alter the spatial pattern of shaking or its impacts. We do AND PLANETARY SCIENCES not consider earthquakes smaller than Mw ∼7.0 because their impacts are likely to be smaller than what are typically considered by contingency planners (for example, the 1988 Dharan and 2011 Sikkim earthquakes, both Mw 6.9), although they may still cause considerable disruption if they affect a major population center. While there is some evidence that earthquakes larger than Mw 8.6, perhaps approaching Mw 9.0, are possible along the Fig. 8. Seismic risk for Nepal. Spatial distribution of relative seismic risk in Himalayan arc (71), this remains contentious (38). Given the Nepal based on summary statistics for modeled fatalities from the ensemble combined with two social vulnerability metrics: (A) percentage of scenarios with scale of potential impacts from Mw 8.6 events compared with at least one fatality, (B) median fatalities for all scenarios that cause fatalities, (C) the extent of Nepal, however, the scale of impacts from an Mw maximum fatalities, (D) specificity of fatalities for all scenarios that cause fatal- 9.0 event may not be substantially larger (SI Appendix, Fig. S2). ities, (E) remoteness score, (F)HDI,and(G) total relative seismic risk, calculated Our scenarios only consider relatively simple fault rupture pat- as the normalized sum of all six risk metrics. Numbers in G show district ranks. terns, ignoring more complex ruptures such as those described by Hamling et al. (72); however, incorporating such complexity into our model requires more advanced seismic modeling, which is SI weightings of the metrics, and so we provide the raw scores in beyond the scope of this study. The potential amplification of Appendix , Tables S2 and S3. ground motion by sedimentary basins, such as the Kathmandu Using our combination, we find that total seismic risk is no- Valley, is also an important factor that has not been included in tably higher in western areas of Nepal (Fig. 8). Gulmi in the this study, along with secondary hazards and cascading hazards West Region is the most at-risk district with a score of 3.61, and such as landsliding and liquefaction. We note, however, that the nearby districts of Rolpa, Pyuthan, and Baglung account for recent improvements in coseismic landslide modeling, including three of the next four most at-risk districts, demonstrating that our ongoing work on this topic (73–75), allow some of these this area has the highest seismic risk in Nepal. Saptari in the East effects to be incorporated into future more holistic iterations of Region has the lowest risk, although many districts on the this approach. Given that the effects of coseismic landslides southern border, particularly in the East and Central Regions, appear to be more pronounced among rural mountainous com- have comparable scores. munities (76), their inclusion may not significantly alter the The middle-to-low score for Kathmandu district (2.61) is general pattern of relative seismic risk established here. particularly notable. While Kathmandu has high scores for the Alternatives to the assignment of uniform weights to all sce- frequency of fatalities and the worst-case scenario, its low spec- narios in the ensemble may also require further exploration. ificity, low remoteness, and comparatively high HDI all help to Herein, we have used a uniform weighting because of gaps in our reduce its total relative seismic risk compared with other districts understanding of earthquake recurrence along the Himalayan arc, in Nepal. It is striking that while Kathmandu commonly features and thus the likelihood of each scenario earthquake in our en- on global rankings of high-seismic-risk cities, ∼9.5 million Ne- semble is unknown. The suggestion that earthquakes of all mag- palis (∼35% of the total population) live in districts with higher nitudes on the MHT may have similar recurrence times (36) may risk scores than the capital, and the 10 most at-risk districts in in part support this assumption. However, while this may be true for Nepal contain a total population of ∼2.5 million, comparable to the MHT, it is unlikely to be so for ruptures of the other upper- the population of the capital. Applying alternate weightings to plate faults included in our ensemble. In locations where recurrence each of the risk metrics changes the values and specific ranking intervals are better constrained, or where Gutenberg–Richter

Robinson et al. PNAS Latest Articles | 7of10 Downloaded by guest on September 24, 2021 relationships are well known, these could be used to derive to predict the precise timing or nature of future earthquakes, and appropriate nonuniform weightings for use in the ensemble. thus their resulting impacts. To date, the assessment of seismic A further limitation relates to assumptions made around short- risk has focused primarily on improving understanding of earth- and long-term population exposure, where basic research could quake hazard in terms of potential ground shaking, which has significantly improve the accuracy of our results. Distributing the resulted in major advances (78). Nevertheless, for contingency population equally between each building type is likely to be an planning, the precise geophysical nature of the earthquake that unrealistic proxy for exposure. Key differences in occupancy are next occurs is of lesser importance than its impacts (25–27). Thus, known between building typologies: reinforced concrete buildings finding an approach that provides insight on what impacts are in Nepal are typically multistory and able to house several families, most likely to happen, and that can complement methods to assess whereas wooden and adobe buildings are smaller and usually only seismic hazard, has obvious benefits. house a single family. The collapse of the former building typology We present an approach to estimating relative seismic risk that therefore likely underestimates impacts, while collapse of other relies on an ensemble of scenarios representing potential future building types may overestimate impacts. earthquakes. This approach is particularly well-suited to coun- Assumptions around the population exposure at different tries like Nepal, where earthquake hazard is relatively poorly times of day are also poorly constrained. Our initial assumptions understood, information on earthquake recurrence intervals is are based on discussions with humanitarian agencies in Nepal limited, and earthquake hazard maps contain widely differing but are likely to be a gross oversimplification. In reality, the results. Our approach weights all plausible future large earth- difference in population exposure between working and non- quakes equally, allowing us to focus on elements of vulnerability working days, particularly in rural areas, may be less pronounced and exposure that contribute to relative seismic risk. Our work than assumed here. Further, the population exposure is likely to shows that it is possible to assess the range of potential impacts be highly spatially variable and not well represented by simple and to consider how specific impacts relate to specific earth- definitions of urban and rural VDCs. We presently lack sufficient quakes. For the majority of districts in Nepal, similar impacts information to fully investigate the effect of temporal variations occur irrespective of the scenario earthquake, and these impacts in exposure; while a simple analysis of night versus day has been are typically closer to the minimum than the worst case. This undertaken, a more nuanced analysis is required to investigate suggests first that the scale of impacts expected in a future how exposure varies diurnally, particularly around communal earthquake can already be relatively well constrained, and sec- times such as meals, and also through the seasons. For instance, ond that planning for the worst-case impacts may place an un- we would expect that population movements change significantly necessarily large burden on the limited resources available. during the monsoon period and during the Tihar and Desai Instead, the optimal level of mitigation that minimizes the total festivals (77), but the effect of these on earthquake risk is yet to cost to society, including both the cost of expected impacts and be addressed. Although we have attempted to calibrate occu- the cost of mitigation (22, 79, 80), may require planning for SI Appendix pancy rates using the 2015 earthquake ( , Fig. S1), losses significantly smaller than the worst case. we note that it is not possible from the available data to de- Our results also imply that, while Kathmandu is regarded as termine whether the departure of the model results relates to one of the most seismically at-risk cities in the world (37, 81), limitations in the occupancy rates, the shaking estimates, the greater relative seismic risk exists in the rural western areas, building fragility curves or, more likely, some combination of particularly in Gulmi and neighboring districts. This suggests these factors. that, while the whole of Nepal requires urgent earthquake risk- While assumptions around population exposure play an im- reduction activities, rural western districts are in particular need. portant role in controlling specific impacts, we highlight that A sole planning focus on urban earthquake risk in Kathmandu these assumptions have been kept consistent throughout our may therefore be inappropriate, as many rural populations ensemble. Thus, while the number of fatalities presented is not within Nepal are at greater relative risk. intended to be absolute, the relative differences between districts should remain unchanged unless there are significant differences ACKNOWLEDGMENTS. The authors thank Rich Walters, Ken McCaffrey, in the movements of people within different districts beyond the Dave Milledge, and Pippa Whitehouse for their helpful comments and urban and rural distinction employed. Limitations associated feedback on early drafts of the manuscript. We thank the editor, David with population exposure serve to further highlight the need for Wald, and an anonymous reviewer for their constructive and detailed reviews, which helped to greatly focus and improve the original manuscript. a more holistic approach to seismic risk analyses. Even if it were This study benefited from the input and assistance of various humanitarian possible to predict the precise timing and nature of a future organizations in Nepal and the wider Asia-Pacific region, including United earthquake, we remain unable to effectively estimate its impacts Nations (UN) Office for Coordination of Humanitarian Affairs Regional if we cannot accurately account for exposure. Office for Asia-Pacific, UN Resident Coordinators Office Nepal, World Food Programme Nepal, Department for International Development, US Agency for International Development, International Federation of the Red Cross Implications and Conclusions and Red Crescent, and European Civil Protection and Humanitarian Aid Advances in our understanding of seismic hazard have long Operations. Funding for this project was provided by the European Union’s shown that for locations such as the Himalayan arc, it is not a Seventh Framework Programme through the DIFeREns 2 COFUND scheme, the Addison Wheeler Fellowship at Durham University, and the Earthquakes matter of whether a devastating earthquake will occur, but when. without Frontiers project (NE/J01995X/1) within the Natural Environmental It is therefore essential to reduce earthquake risk where possible Research Council–Economic and Social Research Council Increasing Resilience and to prepare for this eventuality. We presently remain unable to Natural Hazards Programme.

1. Bilham R (2004) Urban earthquake fatalities: A safer world, or worse to come? Seismol system. Human Casualties in Earthquakes Advances in Natural and Technological Res Lett 75:706–712. Hazards Research, ed Spence R (Springer, New York), pp 83–95. 2. Panza GF, et al. (2011) Advanced seismic hazard assessment. Pure Appl Geophys 168: 6. Jaiswal K, Bausch D, Rozelle J, Holub J, McGowan S (2017) HAZUS estimated annu- 1–9. alized earthquake losses for the United States. Federal Emergency Management 3. Domenico G (1999) The global seismic hazard assessment program (GSHAP)–1992/ Report FEMA P-366 (FEMA, Washington, DC), p 75. 1999. Ann Geophys 42:957–974. 7. Bommer J, et al. (2002) Development of an earthquake loss model for Turkish ca- 4. Jaiswal K, Wald DJ, Hearne M (2009) Estimating casualties for large earthquakes tastrophe insurance. J Seismol 6:431–446. worldwide using an empirical approach, USGS Open File Report 2009-1136 (US Geo- 8. Whitman RV, et al. (1997) Development of a national earthquake loss estimation logical Survey, Reston, VA), p 84. methodology. Earthquake Spectra 13:643–661. 5. Jaiswal K, Wald DJ, Earle PS, Porter KA, Hearne M (2011) Earthquake casualty models 9. Yeh C-H, Loh C-H, Tsai K-C (2006) Overview of Taiwan earthquake loss estimation within the USGS prompt assessment of global earthquakes for response (PAGER) system. Nat Hazards 37:23–37.

8of10 | www.pnas.org/cgi/doi/10.1073/pnas.1807433115 Robinson et al. Downloaded by guest on September 24, 2021 10. Kircher CA, Whitman RV, Holmes WT (2006) HAZUS earthquake loss estimation inversion and their seismotectonic implications. Bull Seismol Soc Am 107: methods. Nat Hazards Rev 7:45–59. 2480–2489. 11. England P, Jackson J (2011) Uncharted seismic risk. Nat Geosci 4:348–349. 44. Sapkota SN, et al. (2012) Primary surface ruptures of the great Himalayan earth- 12. Bilham R (2013) Societal and observational problems in earthquake risk assessments quakes in 1934 and 1255. Nat Geosci 6:71–76. and their delivery to those most at risk. Tectonophysics 584:166–173. 45. Wesnousky SG, et al. (2017) Large paleoearthquake timing and displacement near 13. McGuire RK (1995) Probabilistic seismic hazard analysis and design earthquakes: Damak in eastern Nepal on the Himalayan frontal thrust. Geophys Res Lett 44: Closing the loop. Bull Seismol Soc Am 85:1275–1284. 8219–8226. 14. Bommer JJ (2002) Deterministic vs. probabilistic seismic hazard assessment: An ex- 46. Central Bureau of Statistics (2011) CensusInfo Nepal 2011. Available at dataforall.org/ aggerated and obstructive dichotomy. J Earthquake Eng 6(sup001):43–73. dashboard/nepalcensus/. Accessed January 11, 2018. 15. Stein S, Geller R, Liu M (2011) Bad assumptions or bad luck: Why earthquake hazard 47. Schmidtlein MC, Shafer JM, Berry M, Cutter SL (2011) Modeled earthquake maps need objective testing. Seismol Res Lett 82:623–626. losses and social vulnerability in Charleston, South Carolina. Appl Geogr 31: 16. Wang Z (2011) Seismic hazard assessment: Issues and alternatives. Pure Appl Geophys 269–281. 168:11–25. 48. Patt AG, et al. (2010) Estimating least-developed countries’ vulnerability to 17. Wyss M, Nekrasova A, Kossobokov V (2012) Errors in expected human losses due to climate-related extreme events over the next 50 years. Proc Natl Acad Sci USA incorrect seismic hazard estimates. Nat Hazards 62:927–935. 107:1333–1337. 18. Wang Z, Woolery EW, Shi B, Kiefer JD (2003) Communicating with uncertainty: A 49. Kagan YY, Jackson DD (1994) Long-term probabilistic forecasting of earthquakes. critical issue with probabilistic seismic hazard analysis. Eos Trans Am Geophys Union J Geophys Res Solid Earth 99:13685–13700. 84:501–508. 50. Wells DL, Coppersmith KJ (1994) New empirical relationships among magnitude, 19. Stirling MW (2014) The continued utility of probabilistic seismic-hazard assessment. rupture length, rupture width, rupture area, and surface displacement. Bull Seismol Earthquake Hazard, Risk and Disasters, eds Shroder J, Wyss M (Academic, New York), Soc Am 84:974–1002. – pp 359 376. 51. Mugnier J-L, et al. (2013) Structural interpretation of the great earthquakes of the “ 20. Cramer CH (2004) Comment on Communicating with uncertainty: A critical issue last millennium in the central Himalaya. Earth Sci Rev 127:30–47. ” with probabilistic seismic hazard analysis. Eos Trans Am Geophys Union 85:283. 52. Avouac J-P, Meng L, Wei S, Wang T, Ampuero J-P (2015) Lower edge of locked 21. Musson RMW (2012) PSHA validated by quasi observational means. Seismol Res Lett main Himalayan thrust unzipped by the 2015 Gorkha earthquake. Nat Geosci 8: – 83:130 134. 708–711. 22. Stein S, Brooks EM, Spencer BD, Liu M (2018) Should all of Nepal be treated as having 53. Bollinger L, Tapponnier P, Sapkota SN, Klinger Y (2016) Slip deficit in central the same earthquake hazard? Living Under the Threat of Earthquakes, Springer Nepal: Omen for a repeat of the 1344 AD earthquake? Earth Planets Space Natural Hazards, ed Kruhl JH, Adhikari R, Dorka U (Springer, Cham, Switzerland), pp 68:12. – 27 45. 54. Searle MP (1996) Geological evidence against large-scale pre-holocene offsets along 23. Mulachin L (1996) Development of Caltrans deterministic fault and earthquake haz- the Karakoram fault: Implications for the limited extrustion of the Tibetan Plateau. ard map of California. Eng Geol 42:217–222. Tectonics 15:171–186. 24. Mualchin L (2011) History of modern earthquake hazard mapping and assessment in 55. Brown E, et al. (2002) Slip rates of the Karakorum Fault, Ladakh, India, determined California using a deterministic or scenario approach. Pure Appl Geophys 168: using cosmic ray exposure dating of debris flows and moraines. J Geophys Res Solid 383–407. Earth 107:ESE 7-1–ESE 7-13. 25. International Federation of Red Cross and Red Crescent Societies (2012) Contingency

56. Silver CRP, Murphy MA, Taylor MH, Gosse J, Baltz T (2015) Neotectonics of the EARTH, ATMOSPHERIC, planning guide (International Federation of Red Cross and Red Crescent Societies, Western Nepal Fault System: Implications for Himalayan strain partitioning. Tectonics AND PLANETARY SCIENCES Geneva). Available at www.ifrc.org/PageFiles/40825/1220900-CPG%202012-EN-LR. 34:2494–2513. pdf. Accessed August 15, 2017. 57. Murphy MA, et al. (2014) Limit of strain partitioning in the Himalaya marked by large 26. Inter-Agency Standing Committee (2007) Inter-agency contingency planning guide- earthquakes in western Nepal. Nat Geosci 7:38–42. lines for humanitarian assistance (Inter-Agency Standing Committee, Geneva). 58. Elliott JR, et al. (2010) Extension on the Tibetan plateau: Recent normal Available at www.who.int/hac/network/interagency/ia_guidelines_dec2007.pdf. Ac- faulting measured by InSAR and body wave seismology. Geophys J Int 183: cessed August 15, 2017. 503–535. 27. Inter-Agency Standing Committee (2015) Emergency Response Preparedness 59. Hurtado JM, Hodges KV, Whipple KX (2001) Neotectonics of the Thakkhola graben (ERP) (Inter-Agency Standing Committee, Geneva). Available at https://inter- and implications for recent activity of the South Tibetan fault system in the central agencystandingcommittee.org/system/files/ Nepal Himalaya. Geol Soc Am Bull 113:222–240. emergency_response_preparedness_2015_final.pdf. Accessed August 15, 2017. 60. Field EH, Jordan TH, Cornell CA (2003) OpenSHA: A developing community-modeling 28. Cubasch U, et al. (2001) Projections of future climate changes. Climate Change 2001: environment for seismic hazard analysis. Seismol Res Lett 74:406–419. The Scientific Basis. Contribution of Working Group I to the Third Assessment Report 61. Abrahamson N, Silva W (2008) Summary of the Abrahamson and Silva NGA ground- of the Intergovernmental Panel on Climate Change, eds Houghton JT, et al. (Cam- motion relations. Earthquake Spectra 24:67–97. bridge Univ Press, Cambridge, UK), pp 525–582. 62. Allen TI, Wald DJ (2009) On the use of high-resolution topographic data as a proxy for 29. Murphy JM, et al. (2004) Quantification of modelling uncertainties in a large en- seismic site conditions (Vs30). Bull Seismol Soc Am 99:935–943. semble of climate change simulations. Nature 430:768–772. 63. Wald DJ, Allen TI (2007) Topographic slope as a proxy for seismic site conditions and 30. Kay JE, et al. (2015) The community earth system model (CESM) large ensemble – project: A community resource for studying climate change in the presence of amplification. Bull Seismol Soc Am 97:1379 1395. internal climate variability. Bull Am Meteorol Soc 96:1333–1349. 64. Guragain R (2015) Development of earthquake risk assessment system for Nepal. PhD 31. Fowler HJ, Ekström M (2009) Multi-model ensemble estimates of climate change dissertation (Univ of Tokyo, Tokyo). impacts on UK seasonal precipitation extremes. Int J Climatol 29:385–416. 65. So E (2016) Estimating Fatality Rates for Earthquake Loss Models (Springer, Cham, 32. Moss RH, et al. (2010) The next generation of scenarios for climate change research Switzerland). and assessment. Nature 463:747–756. 66. Gautam D, Fabbrocino G, Santucci de Magistris F (2018) Derive empirical fragility – 33. Zhang Z, Krishnamurti TN, Zhang Z, Krishnamurti TN (1997) Ensemble forecasting of functions for Nepali residential buildings. Eng Struct 171:617 628. hurricane tracks. Bull Am Meteorol Soc 78:2785–2795. 67. USAID-Nepal (2016) Quantifying remoteness. Available at https://www.aiddata.org/ 34. Krishnamurti TN, et al. (2000) Multimodel ensemble forecasts for weather and sea- blog/quantifying-remoteness-a-scale-of-accessibility-across-nepal. Accessed December sonal climate. J Clim 13:4196–4216. 11, 2017. 35. Bollinger L, et al. (2014) Estimating the return times of great Himalayan earthquakes 68. United Nations Development Programme (2014) Nepal human development report in eastern Nepal: Evidence from the patu and Bardibas strands of the main frontal 2014: Beyond geography unlocking human potential. Available at hdr.undp.org/sites/ thrust. J Geophys Res Solid Earth 119:7123–7163. default/files/nepal_nhdr_2014-final.pdf. Accessed August 15, 2017. 36. Feldl N, Bilham R (2006) Great Himalayan earthquakes and the Tibetan plateau. 69. Striessnig E, Lutz W, Patt AG (2013) Effects of educational attainment on climate risk Nature 444:165–170. vulnerability. Ecol Soc 18:16. 37. Bilham R, Gaur VK, Molnar P (2001) Earthquakes. Himalayan seismic hazard. Science 70. United Nations Office for the Coordination of Humanitarian Affairs (2015) Nepal 293:1442–1444. earthquake facts and figures. Available at https://www.undispatch.com/nepal-earth- 38. Bilham R (2005) Apparent Himalayan slip deficit from the summation of seismic quake-facts-and-figures/. Accessed March 5, 2018. > moments for Himalayan earthquakes, 1500–2000. Curr Sci 88:1658–1663. 71. Stevens VL, Avouac J-P (2016) Millenary Mw 9.0 earthquakes required by geodetic 39. United Nations Development Program/Bureau of Crisis Prevention and Recovery strain in the Himalaya. Geophys Res Lett 43:1118–1123. (2004) Reducing disaster risk, a challenge for development (United Nations, New 72. Hamling IJ, et al. (2017) Complex multifault rupture during the 2016 Mw 7.8 Kaikoura¯ York), p 161. earthquake, New Zealand. Science 356:eaam7194. 40. Gautam D (2017) Assessment of social vulnerability to natural hazards in Nepal. Nat 73. Kritikos T, Robinson TR, Davies TRH (2015) Regional coseismic landslide hazard as- Hazards Earth Syst Sci 17:2313–2320. sessment without historical landslide inventories: A new approach. J Geophys Res 41. Wyss M (2017) Four loss estimates for the Gorkha M7.8 earthquake, April 25, 2015, Earth Surf 120:1–19. before and after it occurred. Nat Hazards 86:141–150. 74. Robinson TR, et al. (2017) Rapid post-earthquake modelling of coseismic landslide 42. Chaulagain H, Gautam D, Rodrigues H (2018) Revisiting major historical earthquakes magnitude and distribution for emergency response decision support. Nat Hazards in Nepal: Overview of 1833, 1934, 1980, 1988, 2011 and 2015 seismic events. Impacts Earth Syst Sci 17:1521–1540. and Insights of the Gorkha Earthquake, eds Gautam D, Rodrigues H (Elsevier, Am- 75. Williams JG, et al. (2017) Satellite-based emergency mapping: Landslides triggered by sterdam). the 2015 Nepal earthquake. Nat Hazards Earth Syst Sci 18:185–205. 43. Pettenati F, Sirovich L, Bjerrum LW (2017) Fault sources and kinematics of the 76. Kargel JS, et al. (2016) Geomorphic and geologic controls of geohazards induced by

1897 Mw 8.1 Assam and the 1934 Ms 8.2 Nepal earthquakes retrieved by KF‐NGA Nepal’s 2015 Gorkha earthquake. Science 351:aac8353.

Robinson et al. PNAS Latest Articles | 9of10 Downloaded by guest on September 24, 2021 77. Gill GJ, et al. (2003) Food security and the millennium development goal on hunger in 80. Stein S, Stein J (2013) How good do natural hazard assessments need to be? GSA Asia. Available at https://www.odi.org/sites/odi.org.uk/files/odi-assets/publications- Today 23:60–61. opinion-files/1892.pdf. Accessed March 7, 2018. 81. Elliott JR, et al. (2016) Himalayan megathrust geometry and relation to topography 78. Bendick R, Bilham R (2017) Do weak global stresses synchronize earthquakes? revealed by the Gorkha earthquake. Nat Geosci 9:174–180. Geophys Res Lett 44:8320–8327. 82. Taylor M, Yin A (2009) Active structures of the Himalayan-Tibetan orogen and 79. Stein J, Stein S (2012) Rebuilding Tohoku: A joint geophysical and economic frame- their relationships to earthquake distribution, contemporary strain field, and work for hazard mitigation. GSA Today 22:42–44. Cenozoic volcanism. Geosphere 5:199–214.

10 of 10 | www.pnas.org/cgi/doi/10.1073/pnas.1807433115 Robinson et al. Downloaded by guest on September 24, 2021